🔍 Read the full analysis: Gemini 4 Carbon’s One-Point AI Lead, Explained: What Rankings Leave Out on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Gemini 4 Carbon scored 59 on Artificial Analysis Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. It led four of nine tests but trailed rivals on other tasks, and the benchmark does not establish which model is best for every use. Google has not made Carbon publicly available; it is starting with a limited group of cyber defenders.
Google’s Gemini 4 Carbon has taken the lead on the Artificial Analysis Intelligence Index v5, scoring 59 points to Claude Opus 5.5’s 58 and GPT-6 Astra’s 57. The one-point margin puts Carbon first on this particular benchmark, not across every task: the model led four of the index’s nine tests, lost the other five and is not yet publicly available.
The index, rebuilt by independent benchmarking company Artificial Analysis and launched October 27, combines results from nine tests into one score. Its four equally weighted categories are knowledge work, coding, science and general performance. Carbon’s 59 compares with 50 for Google’s Gemini 4 Argon, its predecessor, as well as 58 for Claude Opus 5.5 and 57 for GPT-6 Astra.
Carbon placed first in tests of difficult reasoning and knowledge questions, scientific work carried out in a terminal, business workflows and knowledge questions that penalize incorrect answers. On the HLE-Diamond test, it scored 69%, versus 68% for Claude Opus 5.5. On Terminal-Bench-Science, it scored 65%, compared with 63% for GPT-6 Astra. The AutomationBench-AA result was 78%, slightly ahead of Argon.
The result is less favorable in several other categories. Claude models led the office-style knowledge-work test, Claude Opus 5.5 scored 46% to Carbon’s 41% on real-world coding, and GPT-6 Astra led the physics-research test. Kimi K3.1 Pro topped terminal tasks, while Meta’s Muse Wonder Full led document reasoning. Artificial Analysis lists Carbon’s cost at $7.76 per index task: less than Claude Opus 5.5’s $10.04, but above GPT-6 Astra’s $6.24 and well above GPT-6.1 Sol’s $1.27.
Infographic 01 · What the #1 is made of
Number one overall. First in four of nine tests.
The Intelligence Index is a decathlon score: nine tests, one number. Carbon wins on reasoning, science and careful knowledge, and trails on office work and coding.
The nine tests behind the score
Carbon at its highest setting (high) against the best model on each test.
| Test | Weight | Carbon | Leader or runner-up |
|---|---|---|---|
| HLE-Diamond Hard reasoning + knowledge questions | Weight 10% | Carbon 69% | Carbon leads Next: Claude Opus 5.5, 68% |
| Terminal-Bench-Science Science research work in a terminal | Weight 15% | Carbon 65% | Carbon leads Next: GPT-6 Astra, 63% |
| AA-Omniscience v1.1 Knowledge, wrong answers penalized | Weight 15% | Carbon 54 | Carbon leads Next: Claude Opus 5.5, 48 |
| AutomationBench-AA Business workflows across apps | Weight 5% | Carbon 78% | Carbon leads Next: Gemini 4 Argon, 78% |
| AA-Briefcase Office-style knowledge work | Weight 15% | Carbon 1,609 Elo | Behind Claude Sonnet 5.5, 1,812 |
| FrontierCode 2.0 Code a maintainer would merge | Weight 15% | Carbon 41% | Behind Claude Opus 5.5, 46% |
| CritPt v2 (hard) Research-level physics | Weight 10% | Carbon 63% | Behind GPT-6 Astra, 71% |
| Terminal-Bench 4.0 Agentic terminal tasks | Weight 10% | Carbon 65% | Behind Kimi K3.1 Pro, 67% |
| GDP.pdf Reasoning over documents | Weight 5% | Carbon 28% | Behind Muse Wonder Full, 33% |
The surprise in the knowledge test
- Accuracy: Claude Fable 5.1 68%, Claude Opus 5.5 66%, GPT-6.1 Sol and GPT-6 Astra 62%, Carbon 58%
- Hallucination rate: Carbon 11%, Claude Opus 5.5 and GPT-6 Astra about 50%
- Carbon wins because it rarely bluffs, not because it knows the most
Read the weights
- Each of the four blocks counts 25%: knowledge work, coding, science, general
- Carbon trails in two of the four heaviest tests (15% each): office-style work and real-world code
- Hallucination rate = wrong answers among all questions not answered correctly
Source: Artificial Analysis, 28 Oct 2026; Index v5 announcement, 27 Oct 2026; Artificial Analysis model data, retrieved 29 Oct 2026. Scores rounded as in the article; Carbon tested at high, other models at their highest tested setting.
Powered by Thorsten Meyer AI
A Narrow Lead With Practical Trade-offs
The one-point difference is a useful signal about performance on the index’s mix of tasks, but it does not settle which model a company or individual should choose. The tests assign equal weight to four broad categories, while particular users may care far more about one task, such as code generation, document analysis or science workflows. Carbon’s losses in coding and office-style knowledge work matter to buyers whose work depends on those skills.
Cost also changes the comparison. The index-task prices supplied by the source show Carbon below Claude Opus 5.5 but above GPT-6 Astra and far above GPT-6.1 Sol. Those figures are not a complete measure of value: a buyer would also need to compare output quality, time, tool use and the expense of correcting errors on its own workload. The ranking is best read as a starting point for testing, not a purchasing instruction.
For everyday users, there is no immediate change because Carbon has not launched for general use. If access expands, the reported strengths suggest its clearest distinction may show up in difficult, long-running tasks rather than routine chat. That is an interpretation of the benchmark results, not a guarantee of performance in consumer products.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
DJI Osmo Pocket 3 Creator Combo Vlogging Camera with 1 CMOS & 4K/120fps
Sonos Era 100 SL Compact, Microphone-Free Speaker – Black
As an affiliate, we earn on qualifying purchases.
How Index v5 Builds Its Score
Artificial Analysis describes its index as a combined score across nine evaluations, with 25% of the total assigned to each of four blocks: knowledge work, coding, science and general performance. The component tests include office-style work, business workflows, coding, terminal tasks, scientific research, hard reasoning and knowledge questions. A model’s overall position can change if the test mix or category weights change.
The company says 60% of v5 uses private test sets or held-out answers, intended to reduce the chance that models have memorized test questions. Many evaluations also involve extended agent tasks, such as working in a terminal, producing reports or changing code. These are more involved than short question-and-answer prompts and contribute to the cost of running the benchmark.
One result needs particular care: Carbon scored 54 on AA-Omniscience, ahead of Claude Opus 5.5 by six points, even though several competitors had higher plain-answer accuracy. The source reports Carbon at 58% accuracy, compared with 66% for Claude Opus 5.5 and 68% for Claude Fable 5.1. The test also penalizes wrong answers; Artificial Analysis reports an 11% hallucination rate for Carbon, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. The index result therefore reflects both answering correctly and avoiding incorrect guesses, not simply how many questions a model gets right.
Access, Pricing and Real-World Results
Google’s release sequence, as reported in the source, starts with a subset of cyber defenders, followed by trusted partners, developers, businesses and consumers. No dates are given for those stages, and the material does not say how many organizations are in the first group or what access conditions apply.
The source lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively. It does not specify when the later rates begin, which customers receive introductory rates or how those token prices translate into total costs for typical workloads. The $7.76 index-task estimate and API token prices measure different things and should not be treated as interchangeable.
The provided material also does not include full results for every index test or detailed information on repeated runs and uncertainty around the one-point margin. It is consequently unclear how stable the ranking is across runs, and whether Carbon’s benchmark advantages will carry over to customer tasks. The source’s discussion of the separate Coding Agent Index is truncated, so it does not provide enough material to report a fuller comparison of coding tools.
Google’s Rollout Sets the Next Test
The next concrete development is whether Google widens Carbon access beyond its initial cyber-defender group and publishes dates for the later stages. Until then, developers and businesses cannot assess the model through general public availability, and consumers have no confirmed launch date.
For organizations evaluating the model, the reported figures point to testing it on representative internal tasks and comparing cost per completed task, not just the index rank. Future results from users, expanded benchmark reporting and details on availability and pricing could show whether Carbon’s high marks on science, reasoning and business workflows outweigh its weaker results in coding and office-style work.
Key Questions
What is Gemini 4 Carbon’s score?
It scored 59 points on Artificial Analysis Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra.
Does the ranking mean Carbon is the best model for every task?
No. Carbon led four of the nine tests and lost the other five. Claude models led office-style knowledge work, Claude Opus 5.5 led the reported real-world coding test, and other models led physics research, terminal tasks and document reasoning.
Why did Carbon lead the knowledge test if other models had higher accuracy?
AA-Omniscience penalizes incorrect answers as well as rewarding correct ones. Artificial Analysis reports that Carbon had lower plain accuracy than several rivals but a lower hallucination rate, which helped it score higher on the test.
Can the public use Gemini 4 Carbon now?
Not according to the supplied report. Google is initially providing it to a subset of cyber defenders, with later access planned for trusted partners, developers, businesses and consumers. No dates for those stages were provided.
What does the reported API pricing mean?
The source lists introductory rates of $2 per million input tokens and $10 per million output tokens, followed later by $4 and $20. It does not specify when the later rates begin or provide enough information to estimate total costs for a particular workload.
Source: ThorstenMeyerAI.com
